What have we Learned from Structural Models?
Learning From Data in Economics
Richard Blundell (UCL and IFS)
Draft paper and full set of references on AEA website
(also my webpage http://www.ucl.ac.uk/ uctp39a/)
AEA, Chicago, January 2017
Richard Blundell (UCL and IFS)
Structural Models: What have we learned?
AEA, Chicago, January 2017
1 / 12
Learning from Data and Structural Models
Focus on structural (micro)econometric models for policy analysis.
Broad definition: A structural economic model is one where the structure
of decision making is (fully) incorporated in the specification of the model.
Richard Blundell (UCL and IFS)
Structural Models: What have we learned?
AEA, Chicago, January 2017
2 / 12
Learning from Data and Structural Models
Focus on structural (micro)econometric models for policy analysis.
Broad definition: A structural economic model is one where the structure
of decision making is (fully) incorporated in the specification of the model.
Aim to identify:
1
structural ‘deep’ parameters: e.g. Frisch vs Marshallian elasticities.
2
underlying mechanisms: e.g. partial vs self-insurance.
3
policy counterfactuals: e.g. ex-ante tax policy evaluation.
Richard Blundell (UCL and IFS)
Structural Models: What have we learned?
AEA, Chicago, January 2017
2 / 12
Learning from Data and Structural Models
Focus on structural (micro)econometric models for policy analysis.
Broad definition: A structural economic model is one where the structure
of decision making is (fully) incorporated in the specification of the model.
Aim to identify:
1
structural ‘deep’ parameters: e.g. Frisch vs Marshallian elasticities.
2
underlying mechanisms: e.g. partial vs self-insurance.
3
policy counterfactuals: e.g. ex-ante tax policy evaluation.
The ability to provide counterfactual policy simulations set structural
models apart from ‘reduced-form’/treatment effect approaches.
But require the detailed specification of the decision-making problem.
Typically placing tougher (explicit) conditions on measurement and
relying, in part, on stronger identifying assumptions.
Richard Blundell (UCL and IFS)
Structural Models: What have we learned?
AEA, Chicago, January 2017
2 / 12
Learning from Data and Structural Models
Focus on structural (micro)econometric models for policy analysis.
Broad definition: A structural economic model is one where the structure
of decision making is (fully) incorporated in the specification of the model.
Aim to identify:
1
structural ‘deep’ parameters: e.g. Frisch vs Marshallian elasticities.
2
underlying mechanisms: e.g. partial vs self-insurance.
3
policy counterfactuals: e.g. ex-ante tax policy evaluation.
The ability to provide counterfactual policy simulations set structural
models apart from ‘reduced-form’/treatment effect approaches.
But require the detailed specification of the decision-making problem.
Typically placing tougher (explicit) conditions on measurement and
relying, in part, on stronger identifying assumptions.
Distinguish identification of full-structural from quasi-structural models.
Richard Blundell (UCL and IFS)
Structural Models: What have we learned?
AEA, Chicago, January 2017
2 / 12
Draw on three related areas in empirical micro/econometrics:
1
2
3
Revealed preference and unobserved heterogeneity.
Discrete choice and welfare reform.
Dynamic structural models with human capital investments.
Richard Blundell (UCL and IFS)
Structural Models: What have we learned?
AEA, Chicago, January 2017
3 / 12
Draw on three related areas in empirical micro/econometrics:
1
2
3
Revealed preference and unobserved heterogeneity.
Discrete choice and welfare reform.
Dynamic structural models with human capital investments.
These areas have extensive policy applications:
where structural functions and policy counterfactuals are well defined.
that ask well-formulated questions e.g. from the ex-post impact of
taxes, prices & wages, to ex-ante predictions and optimal design;
analyse static choice and dynamic choice;
use new data, new methods and new computational developments.
Richard Blundell (UCL and IFS)
Structural Models: What have we learned?
AEA, Chicago, January 2017
3 / 12
Draw on three related areas in empirical micro/econometrics:
1
2
3
Revealed preference and unobserved heterogeneity.
Discrete choice and welfare reform.
Dynamic structural models with human capital investments.
These areas have extensive policy applications:
where structural functions and policy counterfactuals are well defined.
that ask well-formulated questions e.g. from the ex-post impact of
taxes, prices & wages, to ex-ante predictions and optimal design;
analyse static choice and dynamic choice;
use new data, new methods and new computational developments.
I emphasize:
a. Models that minimize assumptions on the structural function and on
unobserved heterogeneity.
b. Approaches that align structural and ‘reduced form’ moments/
treatment effects.
Richard Blundell (UCL and IFS)
Structural Models: What have we learned?
AEA, Chicago, January 2017
3 / 12
Draw on three related areas in empirical micro/econometrics:
1
2
3
Revealed preference and unobserved heterogeneity.
Discrete choice and welfare reform.
Dynamic structural models with human capital investments.
These areas have extensive policy applications:
where structural functions and policy counterfactuals are well defined.
that ask well-formulated questions e.g. from the ex-post impact of
taxes, prices & wages, to ex-ante predictions and optimal design;
analyse static choice and dynamic choice;
use new data, new methods and new computational developments.
I emphasize:
a. Models that minimize assumptions on the structural function and on
unobserved heterogeneity.
b. Approaches that align structural and ‘reduced form’ moments/
treatment effects.
Mainly single agent models: do not focus here on equilibrium concepts,
incompleteness and coherency: IO, networks, ...
Richard Blundell (UCL and IFS)
Structural Models: What have we learned?
AEA, Chicago, January 2017
3 / 12
1. Revealed Preference with Heterogeneity
Consider a simple structural demand function of interest:
y = g (p, I , z, u )
which describes demand for good(s) ‘y ’ by consumer (z, u ) facing (p, I ).
RP inequalities summarise structural information:
• Use RP to bound (set identify) individual demand counterfactuals,
- if reject, analyse taste change /derive inequalities for alternative theories.
Richard Blundell (UCL and IFS)
Structural Models: What have we learned?
AEA, Chicago, January 2017
4 / 12
1. Revealed Preference with Heterogeneity
Consider a simple structural demand function of interest:
y = g (p, I , z, u )
which describes demand for good(s) ‘y ’ by consumer (z, u ) facing (p, I ).
RP inequalities summarise structural information:
• Use RP to bound (set identify) individual demand counterfactuals,
- if reject, analyse taste change /derive inequalities for alternative theories.
∗ additive u is implausible - with nonseparable u & monotonicity,
conditional nonparametric quantiles identify individual demands.
∗ relax conditional independence ass. on u to allow endogeneity of prices p.
∗ with multivariate heterogeneity only identify average structural welfare.
∗ with multiple demands invertibility in u is not sufficient for identification.
Richard Blundell (UCL and IFS)
Structural Models: What have we learned?
AEA, Chicago, January 2017
4 / 12
1. Revealed Preference with Heterogeneity
Consider a simple structural demand function of interest:
y = g (p, I , z, u )
which describes demand for good(s) ‘y ’ by consumer (z, u ) facing (p, I ).
RP inequalities summarise structural information:
• Use RP to bound (set identify) individual demand counterfactuals,
- if reject, analyse taste change /derive inequalities for alternative theories.
∗ additive u is implausible - with nonseparable u & monotonicity,
conditional nonparametric quantiles identify individual demands.
∗ relax conditional independence ass. on u to allow endogeneity of prices p.
∗ with multivariate heterogeneity only identify average structural welfare.
∗ with multiple demands invertibility in u is not sufficient for identification.
New methods address multiple demands and multiple heterogeneity using
proxies/‘excluded covariates’ with ‘large’ data to identify counterfactuals.
Richard Blundell (UCL and IFS)
Structural Models: What have we learned?
AEA, Chicago, January 2017
4 / 12
1. Revealed Preference with Heterogeneity
Consider a simple structural demand function of interest:
y = g (p, I , z, u )
which describes demand for good(s) ‘y ’ by consumer (z, u ) facing (p, I ).
RP inequalities summarise structural information:
• Use RP to bound (set identify) individual demand counterfactuals,
- if reject, analyse taste change /derive inequalities for alternative theories.
∗ additive u is implausible - with nonseparable u & monotonicity,
conditional nonparametric quantiles identify individual demands.
∗ relax conditional independence ass. on u to allow endogeneity of prices p.
∗ with multivariate heterogeneity only identify average structural welfare.
∗ with multiple demands invertibility in u is not sufficient for identification.
New methods address multiple demands and multiple heterogeneity using
proxies/‘excluded covariates’ with ‘large’ data to identify counterfactuals.
=> New insights on consumer preferences - e.g. gasoline demand...
Richard Blundell (UCL and IFS)
Structural Models: What have we learned?
AEA, Chicago, January 2017
4 / 12
Gasoline example: quantile demands under RP inequalities:
Symmetrical joint confidence intervals (50−th quantile), middle income group
CI
CI
unconstrained exog
constrained exog
7.8
7.6
log demand
7.4
7.2
7
6.8
6.6
0.2
0.22
0.24
0.26
0.28
log price
0.3
0.32
0.34
0.36
Source: Blundell, Horowtiz and Parey (2015)
* price/tax responses are non-monotonic across the income distribution,
* calculate distribution of individual welfare costs of gasoline tax.
Richard Blundell (UCL and IFS)
Structural Models: What have we learned?
AEA, Chicago, January 2017
5 / 12
2. Structural discrete choice models for policy analysis
Use labor supply and welfare-benefit reform to highlight ‘five key steps’
involved in learning from data in empirical policy design, (Mirrlees Review)
1
Key margins of adjustment
2
Measurement of effective incentives
3
The importance of information and complexity
4
Evidence on the size of responses and on causality
5
Implications for policy design
Richard Blundell (UCL and IFS)
Structural Models: What have we learned?
AEA, Chicago, January 2017
6 / 12
2. Structural discrete choice models for policy analysis
Use labor supply and welfare-benefit reform to highlight ‘five key steps’
involved in learning from data in empirical policy design, (Mirrlees Review)
1
Key margins of adjustment
2
Measurement of effective incentives
3
The importance of information and complexity
4
Evidence on the size of responses and on causality
5
Implications for policy design
-> Step 4: quasi-experimental approaches can provide robust measures of
ex-post policy impacts but are necessarily local and limited in scope.
-> Structural models uncover mechanisms and provide counterfactual
policy simulations which feed into a policy (re-)design in step 5.
Richard Blundell (UCL and IFS)
Structural Models: What have we learned?
AEA, Chicago, January 2017
6 / 12
Key elements of a structural labor supply model for welfare reform:
1
precise definition of budget constraint - tax and benefit interactions.
2
specification of preferences for discrete hours & benefit combinations.
3
essential heterogeneity in ‘tastes’ for work, stigma and fixed costs.
4
identification using variation in welfare/tax rules by time and location.
Richard Blundell (UCL and IFS)
Structural Models: What have we learned?
AEA, Chicago, January 2017
7 / 12
Key elements of a structural labor supply model for welfare reform:
1
precise definition of budget constraint - tax and benefit interactions.
2
specification of preferences for discrete hours & benefit combinations.
3
essential heterogeneity in ‘tastes’ for work, stigma and fixed costs.
4
identification using variation in welfare/tax rules by time and location.
Findings:
∗ Ex-ante counterfactuals match post-reform behavior, (see references)
∗ Extensive responses larger for low educated mothers with young children.
∗ Derive optimal structure for tax-credit design -> UK policy impact.
Richard Blundell (UCL and IFS)
Structural Models: What have we learned?
AEA, Chicago, January 2017
7 / 12
Key elements of a structural labor supply model for welfare reform:
1
precise definition of budget constraint - tax and benefit interactions.
2
specification of preferences for discrete hours & benefit combinations.
3
essential heterogeneity in ‘tastes’ for work, stigma and fixed costs.
4
identification using variation in welfare/tax rules by time and location.
Findings:
∗ Ex-ante counterfactuals match post-reform behavior, (see references)
∗ Extensive responses larger for low educated mothers with young children.
∗ Derive optimal structure for tax-credit design -> UK policy impact.
Can we use stochastic nonparametric RP?
∗ Adapt RP inequalities to discrete outcomes and nonlinear budget sets:
∗ Conditions on heterogeneity for partial identification of counterfactuals,
∗ Recent applications identify quasi-structural contrasts - taxable income
functions and derivatives (taxable income elasticities).
Richard Blundell (UCL and IFS)
Structural Models: What have we learned?
AEA, Chicago, January 2017
7 / 12
3. Structural Dynamic Models
Life-cycle labor supply and human capital investment in tax policy/design.
∗ Identification of structural dynamic models requires strong assumptions
-> on subjective discount rates and distribution of beliefs.
=> quasi-structural models are more valuable.
∗ Identify key life-cycle counterfactuals and mechanisms
-> e.g. ‘insurance value’ of redistributive policies - essential for tax design.
∗ Human capital investment breaks intertemporal separability
-> e.g. wage depends on experience capital in learning-by-doing model.
=>quasi-structural models and quasi-experimental contrasts more difficult.
Richard Blundell (UCL and IFS)
Structural Models: What have we learned?
AEA, Chicago, January 2017
8 / 12
3. Structural Dynamic Models
Life-cycle labor supply and human capital investment in tax policy/design.
∗ Identification of structural dynamic models requires strong assumptions
-> on subjective discount rates and distribution of beliefs.
=> quasi-structural models are more valuable.
∗ Identify key life-cycle counterfactuals and mechanisms
-> e.g. ‘insurance value’ of redistributive policies - essential for tax design.
∗ Human capital investment breaks intertemporal separability
-> e.g. wage depends on experience capital in learning-by-doing model.
=>quasi-structural models and quasi-experimental contrasts more difficult.
• Draw on a specific application: structural model of education choices,
human capital and labor supply in the UK: use linked admin/panel data
and the time series of local tax, tax-credit, welfare, and tuition reforms to
identify parameters, conditioning on early life-history variables.
Richard Blundell (UCL and IFS)
Structural Models: What have we learned?
AEA, Chicago, January 2017
8 / 12
3. Structural Dynamic Models
Life-cycle labor supply and human capital investment in tax policy/design.
∗ Identification of structural dynamic models requires strong assumptions
-> on subjective discount rates and distribution of beliefs.
=> quasi-structural models are more valuable.
∗ Identify key life-cycle counterfactuals and mechanisms
-> e.g. ‘insurance value’ of redistributive policies - essential for tax design.
∗ Human capital investment breaks intertemporal separability
-> e.g. wage depends on experience capital in learning-by-doing model.
=>quasi-structural models and quasi-experimental contrasts more difficult.
• Draw on a specific application: structural model of education choices,
human capital and labor supply in the UK: use linked admin/panel data
and the time series of local tax, tax-credit, welfare, and tuition reforms to
identify parameters, conditioning on early life-history variables.
For example, at the heart is a dynamic structural wage function estimated
jointly with education and life-cycle labor supply decisions =>
Richard Blundell (UCL and IFS)
Structural Models: What have we learned?
AEA, Chicago, January 2017
8 / 12
Dynamic structural wage function (women):
1.8
2
log wage
2.2
2.4
2.6
2.8
ln wsit =ln Wsit + γsi ln (esit + 1) + υsit + ξ sit
20
30
40
50
age
Sec
HS
Univ
Figure: Implied female wage profiles from structural model and the data (by education).
Source: Blundell, Dias, Meghir and Shaw (2016)
Richard Blundell (UCL and IFS)
Structural Models: What have we learned?
AEA, Chicago, January 2017
9 / 12
Dynamic structural wage function (women):
1.8
2
log wage
2.2
2.4
2.6
2.8
ln wsit =ln Wsit + γsi ln (esit + 1) + υsit + ξ sit
20
30
40
50
age
Sec
HS
Univ
Figure: Implied female wage profiles from structural model and the data (by education).
Source: Blundell, Dias, Meghir and Shaw (2016)
* Structural model fit is good (also for labor supply and education choices)
-> only achieved by allowing the experience returns to differ by education,
-> conditioning on extensive background factors,
-> and including a part-time work penalty in experience capital.
Richard Blundell (UCL and IFS)
Structural Models: What have we learned?
AEA, Chicago, January 2017
9 / 12
(Some) lessons learned
∗ Structural parameters:
- Experience effects display strong dynamic complementarity, with lower
experience effects for low educated and those in part-time work.
∗ Mechanisms:
- The insurance value of tax-credits is substantial part of welfare gain.
∗ Counterfactuals and policy design:
- Lower education women with young kids have larger supply responses
=> tax credits are an optimal design (but little earnings progression).
- Significant, but small, effect of tax credits on education choice,
=> attenuating some of the employment gains.
• Reconciliation of past results:
- Can explain past (static) structural and quasi-experimental results human capital depreciation and part-time work imply negligible experience
capital wage dynamics/non-separability for low educated women.
Richard Blundell (UCL and IFS)
Structural Models: What have we learned?
AEA, Chicago, January 2017
10 / 12
Quasi-Structural Models
Identify a subset of structural parameters and/or mechanisms rather than
full counterfactuals. Can also be used to test/assess the full model spec.
Three examples from dynamic structural models (references in paper):
Richard Blundell (UCL and IFS)
Structural Models: What have we learned?
AEA, Chicago, January 2017
11 / 12
Quasi-Structural Models
Identify a subset of structural parameters and/or mechanisms rather than
full counterfactuals. Can also be used to test/assess the full model spec.
Three examples from dynamic structural models (references in paper):
1
under intertemporal separability, ‘life-cycle’ consistent within period
preferences are identified by conditioning on consumption (net saving)
which represents a sufficient static for future expectations and
discount rates, and robust to liquidity constraints
2
partial insurance, in which structural ’insurance’ parameters are
identified without parametric specification of expectations. Can also
test for advanced information.
3
estimation of an Euler equation, which identifies ’Frisch’ elasticities
under weaker conditions on information and expectations
• Do not identify full life-cycle counterfactuals.
• Note recent RCTs designed to identify subsets of structural parameters.
Richard Blundell (UCL and IFS)
Structural Models: What have we learned?
AEA, Chicago, January 2017
11 / 12
Summary: Structural econometric models can deliver new insights
(i) the recovery of structural ‘deep’ parameters: e.g. the size of extensive
and intensive elasticities in a dynamic model with experience capital.
(ii) the identification of underlying mechanisms: e.g. the separation of
incentive effects from insurance effects in tax policy reform.
(iii) counterfactuals and policy design: e.g. the ex-ante simulation of a
reform to the tax-credit and welfare system.
But full counterfactuals require careful modelling of preferences,
constraints, non-separabilities and heterogeneity.
Quasi-structural approaches can relax some of these modeling assumptions
and quasi-experimental approaches can deliver robust moments to match.
Richard Blundell (UCL and IFS)
Structural Models: What have we learned?
AEA, Chicago, January 2017
12 / 12
Summary: Structural econometric models can deliver new insights
(i) the recovery of structural ‘deep’ parameters: e.g. the size of extensive
and intensive elasticities in a dynamic model with experience capital.
(ii) the identification of underlying mechanisms: e.g. the separation of
incentive effects from insurance effects in tax policy reform.
(iii) counterfactuals and policy design: e.g. the ex-ante simulation of a
reform to the tax-credit and welfare system.
But full counterfactuals require careful modelling of preferences,
constraints, non-separabilities and heterogeneity.
Quasi-structural approaches can relax some of these modeling assumptions
and quasi-experimental approaches can deliver robust moments to match.
The last words to Ragnar Frisch...... “No amount of statistical information,
however complete and exact, can by itself explain economic phenomena. If
we are not to get lost in the overwhelming, bewildering mass of statistical
data, we need the guidance of a powerful theoretical framework. Without
this no significant interpretation of our observations will be possible.”
Richard Blundell (UCL and IFS)
Structural Models: What have we learned?
AEA, Chicago, January 2017
12 / 12